mm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  mm/userfaultfd.c
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched/signal.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/rmap.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/shmem_fs.h>
  18 #include <asm/tlbflush.h>
  19 #include <asm/tlb.h>
  20 #include "internal.h"
  21
  22 static __always_inline
  23 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  24                                     unsigned long dst_start,
  25                                     unsigned long len)
  26 {
  27         /*
  28          * Make sure that the dst range is both valid and fully within a
  29          * single existing vma.
  30          */
  31         struct vm_area_struct *dst_vma;
  32
  33         dst_vma = find_vma(dst_mm, dst_start);
  34         if (!dst_vma)
  35                 return NULL;
  36
  37         if (dst_start < dst_vma->vm_start ||
  38             dst_start + len > dst_vma->vm_end)
  39                 return NULL;
  40
  41         /*
  42          * Check the vma is registered in uffd, this is required to
  43          * enforce the VM_MAYWRITE check done at uffd registration
  44          * time.
  45          */
  46         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  47                 return NULL;
  48
  49         return dst_vma;
  50 }
  51
  52 /*
  53  * Install PTEs, to map dst_addr (within dst_vma) to page.
  54  *
  55  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
  56  * and anon, and for both shared and private VMAs.
  57  */
  58 int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
  59                              struct vm_area_struct *dst_vma,
  60                              unsigned long dst_addr, struct page *page,
  61                              bool newly_allocated, bool wp_copy)
  62 {
  63         int ret;
  64         pte_t _dst_pte, *dst_pte;
  65         bool writable = dst_vma->vm_flags & VM_WRITE;
  66         bool vm_shared = dst_vma->vm_flags & VM_SHARED;
  67         bool page_in_cache = page->mapping;
  68         spinlock_t *ptl;
  69         struct inode *inode;
  70         pgoff_t offset, max_off;
  71
  72         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  73         _dst_pte = pte_mkdirty(_dst_pte);
  74         if (page_in_cache && !vm_shared)
  75                 writable = false;
  76
  77         /*
  78          * Always mark a PTE as write-protected when needed, regardless of
  79          * VM_WRITE, which the user might change.
  80          */
  81         if (wp_copy) {
  82                 _dst_pte = pte_mkuffd_wp(_dst_pte);
  83                 writable = false;
  84         }
  85
  86         if (writable)
  87                 _dst_pte = pte_mkwrite(_dst_pte);
  88         else
  89                 /*
  90                  * We need this to make sure write bit removed; as mk_pte()
  91                  * could return a pte with write bit set.
  92                  */
  93                 _dst_pte = pte_wrprotect(_dst_pte);
  94
  95         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  96
  97         if (vma_is_shmem(dst_vma)) {
  98                 /* serialize against truncate with the page table lock */
  99                 inode = dst_vma->vm_file->f_inode;
 100                 offset = linear_page_index(dst_vma, dst_addr);
 101                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 102                 ret = -EFAULT;
 103                 if (unlikely(offset >= max_off))
 104                         goto out_unlock;
 105         }
 106
 107         ret = -EEXIST;
 108         /*
 109          * We allow to overwrite a pte marker: consider when both MISSING|WP
 110          * registered, we firstly wr-protect a none pte which has no page cache
 111          * page backing it, then access the page.
 112          */
 113         if (!pte_none_mostly(*dst_pte))
 114                 goto out_unlock;
 115
 116         if (page_in_cache) {
 117                 /* Usually, cache pages are already added to LRU */
 118                 if (newly_allocated)
 119                         lru_cache_add(page);
 120                 page_add_file_rmap(page, dst_vma, false);
 121         } else {
 122                 page_add_new_anon_rmap(page, dst_vma, dst_addr);
 123                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 124         }
 125
 126         /*
 127          * Must happen after rmap, as mm_counter() checks mapping (via
 128          * PageAnon()), which is set by __page_set_anon_rmap().
 129          */
 130         inc_mm_counter(dst_mm, mm_counter(page));
 131
 132         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 133
 134         /* No need to invalidate - it was non-present before */
 135         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 136         ret = 0;
 137 out_unlock:
 138         pte_unmap_unlock(dst_pte, ptl);
 139         return ret;
 140 }
 141
 142 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 143                             pmd_t *dst_pmd,
 144                             struct vm_area_struct *dst_vma,
 145                             unsigned long dst_addr,
 146                             unsigned long src_addr,
 147                             struct page **pagep,
 148                             bool wp_copy)
 149 {
 150         void *page_kaddr;
 151         int ret;
 152         struct page *page;
 153
 154         if (!*pagep) {
 155                 ret = -ENOMEM;
 156                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 157                 if (!page)
 158                         goto out;
 159
 160                 page_kaddr = kmap_atomic(page);
 161                 ret = copy_from_user(page_kaddr,
 162                                      (const void __user *) src_addr,
 163                                      PAGE_SIZE);
 164                 kunmap_atomic(page_kaddr);
 165
 166                 /* fallback to copy_from_user outside mmap_lock */
 167                 if (unlikely(ret)) {
 168                         ret = -ENOENT;
 169                         *pagep = page;
 170                         /* don't free the page */
 171                         goto out;
 172                 }
 173
 174                 flush_dcache_page(page);
 175         } else {
 176                 page = *pagep;
 177                 *pagep = NULL;
 178         }
 179
 180         /*
 181          * The memory barrier inside __SetPageUptodate makes sure that
 182          * preceding stores to the page contents become visible before
 183          * the set_pte_at() write.
 184          */
 185         __SetPageUptodate(page);
 186
 187         ret = -ENOMEM;
 188         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 189                 goto out_release;
 190
 191         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 192                                        page, true, wp_copy);
 193         if (ret)
 194                 goto out_release;
 195 out:
 196         return ret;
 197 out_release:
 198         put_page(page);
 199         goto out;
 200 }
 201
 202 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 203                               pmd_t *dst_pmd,
 204                               struct vm_area_struct *dst_vma,
 205                               unsigned long dst_addr)
 206 {
 207         pte_t _dst_pte, *dst_pte;
 208         spinlock_t *ptl;
 209         int ret;
 210         pgoff_t offset, max_off;
 211         struct inode *inode;
 212
 213         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 214                                          dst_vma->vm_page_prot));
 215         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 216         if (dst_vma->vm_file) {
 217                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 218                 inode = dst_vma->vm_file->f_inode;
 219                 offset = linear_page_index(dst_vma, dst_addr);
 220                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 221                 ret = -EFAULT;
 222                 if (unlikely(offset >= max_off))
 223                         goto out_unlock;
 224         }
 225         ret = -EEXIST;
 226         if (!pte_none(*dst_pte))
 227                 goto out_unlock;
 228         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 229         /* No need to invalidate - it was non-present before */
 230         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 231         ret = 0;
 232 out_unlock:
 233         pte_unmap_unlock(dst_pte, ptl);
 234         return ret;
 235 }
 236
 237 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 238 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 239                                 pmd_t *dst_pmd,
 240                                 struct vm_area_struct *dst_vma,
 241                                 unsigned long dst_addr,
 242                                 bool wp_copy)
 243 {
 244         struct inode *inode = file_inode(dst_vma->vm_file);
 245         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 246         struct folio *folio;
 247         struct page *page;
 248         int ret;
 249
 250         ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
 251         /* Our caller expects us to return -EFAULT if we failed to find folio */
 252         if (ret == -ENOENT)
 253                 ret = -EFAULT;
 254         if (ret)
 255                 goto out;
 256         if (!folio) {
 257                 ret = -EFAULT;
 258                 goto out;
 259         }
 260
 261         page = folio_file_page(folio, pgoff);
 262         if (PageHWPoison(page)) {
 263                 ret = -EIO;
 264                 goto out_release;
 265         }
 266
 267         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 268                                        page, false, wp_copy);
 269         if (ret)
 270                 goto out_release;
 271
 272         folio_unlock(folio);
 273         ret = 0;
 274 out:
 275         return ret;
 276 out_release:
 277         folio_unlock(folio);
 278         folio_put(folio);
 279         goto out;
 280 }
 281
 282 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 283 {
 284         pgd_t *pgd;
 285         p4d_t *p4d;
 286         pud_t *pud;
 287
 288         pgd = pgd_offset(mm, address);
 289         p4d = p4d_alloc(mm, pgd, address);
 290         if (!p4d)
 291                 return NULL;
 292         pud = pud_alloc(mm, p4d, address);
 293         if (!pud)
 294                 return NULL;
 295         /*
 296          * Note that we didn't run this because the pmd was
 297          * missing, the *pmd may be already established and in
 298          * turn it may also be a trans_huge_pmd.
 299          */
 300         return pmd_alloc(mm, pud, address);
 301 }
 302
 303 #ifdef CONFIG_HUGETLB_PAGE
 304 /*
 305  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 306  * called with mmap_lock held, it will release mmap_lock before returning.
 307  */
 308 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 309                                               struct vm_area_struct *dst_vma,
 310                                               unsigned long dst_start,
 311                                               unsigned long src_start,
 312                                               unsigned long len,
 313                                               enum mcopy_atomic_mode mode,
 314                                               bool wp_copy)
 315 {
 316         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 317         ssize_t err;
 318         pte_t *dst_pte;
 319         unsigned long src_addr, dst_addr;
 320         long copied;
 321         struct page *page;
 322         unsigned long vma_hpagesize;
 323         pgoff_t idx;
 324         u32 hash;
 325         struct address_space *mapping;
 326
 327         /*
 328          * There is no default zero huge page for all huge page sizes as
 329          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 330          * by THP.  Since we can not reliably insert a zero page, this
 331          * feature is not supported.
 332          */
 333         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 334                 mmap_read_unlock(dst_mm);
 335                 return -EINVAL;
 336         }
 337
 338         src_addr = src_start;
 339         dst_addr = dst_start;
 340         copied = 0;
 341         page = NULL;
 342         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 343
 344         /*
 345          * Validate alignment based on huge page size
 346          */
 347         err = -EINVAL;
 348         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 349                 goto out_unlock;
 350
 351 retry:
 352         /*
 353          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 354          * retry, dst_vma will be set to NULL and we must lookup again.
 355          */
 356         if (!dst_vma) {
 357                 err = -ENOENT;
 358                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 359                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 360                         goto out_unlock;
 361
 362                 err = -EINVAL;
 363                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 364                         goto out_unlock;
 365
 366                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 367         }
 368
 369         /*
 370          * If not shared, ensure the dst_vma has a anon_vma.
 371          */
 372         err = -ENOMEM;
 373         if (!vm_shared) {
 374                 if (unlikely(anon_vma_prepare(dst_vma)))
 375                         goto out_unlock;
 376         }
 377
 378         while (src_addr < src_start + len) {
 379                 BUG_ON(dst_addr >= dst_start + len);
 380
 381                 /*
 382                  * Serialize via vma_lock and hugetlb_fault_mutex.
 383                  * vma_lock ensures the dst_pte remains valid even
 384                  * in the case of shared pmds.  fault mutex prevents
 385                  * races with other faulting threads.
 386                  */
 387                 idx = linear_page_index(dst_vma, dst_addr);
 388                 mapping = dst_vma->vm_file->f_mapping;
 389                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 390                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 391                 hugetlb_vma_lock_read(dst_vma);
 392
 393                 err = -ENOMEM;
 394                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 395                 if (!dst_pte) {
 396                         hugetlb_vma_unlock_read(dst_vma);
 397                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 398                         goto out_unlock;
 399                 }
 400
 401                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 402                     !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
 403                         err = -EEXIST;
 404                         hugetlb_vma_unlock_read(dst_vma);
 405                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 406                         goto out_unlock;
 407                 }
 408
 409                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 410                                                dst_addr, src_addr, mode, &page,
 411                                                wp_copy);
 412
 413                 hugetlb_vma_unlock_read(dst_vma);
 414                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 415
 416                 cond_resched();
 417
 418                 if (unlikely(err == -ENOENT)) {
 419                         mmap_read_unlock(dst_mm);
 420                         BUG_ON(!page);
 421
 422                         err = copy_huge_page_from_user(page,
 423                                                 (const void __user *)src_addr,
 424                                                 vma_hpagesize / PAGE_SIZE,
 425                                                 true);
 426                         if (unlikely(err)) {
 427                                 err = -EFAULT;
 428                                 goto out;
 429                         }
 430                         mmap_read_lock(dst_mm);
 431
 432                         dst_vma = NULL;
 433                         goto retry;
 434                 } else
 435                         BUG_ON(page);
 436
 437                 if (!err) {
 438                         dst_addr += vma_hpagesize;
 439                         src_addr += vma_hpagesize;
 440                         copied += vma_hpagesize;
 441
 442                         if (fatal_signal_pending(current))
 443                                 err = -EINTR;
 444                 }
 445                 if (err)
 446                         break;
 447         }
 448
 449 out_unlock:
 450         mmap_read_unlock(dst_mm);
 451 out:
 452         if (page)
 453                 put_page(page);
 454         BUG_ON(copied < 0);
 455         BUG_ON(err > 0);
 456         BUG_ON(!copied && !err);
 457         return copied ? copied : err;
 458 }
 459 #else /* !CONFIG_HUGETLB_PAGE */
 460 /* fail at build time if gcc attempts to use this */
 461 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 462                                       struct vm_area_struct *dst_vma,
 463                                       unsigned long dst_start,
 464                                       unsigned long src_start,
 465                                       unsigned long len,
 466                                       enum mcopy_atomic_mode mode,
 467                                       bool wp_copy);
 468 #endif /* CONFIG_HUGETLB_PAGE */
 469
 470 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 471                                                 pmd_t *dst_pmd,
 472                                                 struct vm_area_struct *dst_vma,
 473                                                 unsigned long dst_addr,
 474                                                 unsigned long src_addr,
 475                                                 struct page **page,
 476                                                 enum mcopy_atomic_mode mode,
 477                                                 bool wp_copy)
 478 {
 479         ssize_t err;
 480
 481         if (mode == MCOPY_ATOMIC_CONTINUE) {
 482                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 483                                             wp_copy);
 484         }
 485
 486         /*
 487          * The normal page fault path for a shmem will invoke the
 488          * fault, fill the hole in the file and COW it right away. The
 489          * result generates plain anonymous memory. So when we are
 490          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 491          * generate anonymous memory directly without actually filling
 492          * the hole. For the MAP_PRIVATE case the robustness check
 493          * only happens in the pagetable (to verify it's still none)
 494          * and not in the radix tree.
 495          */
 496         if (!(dst_vma->vm_flags & VM_SHARED)) {
 497                 if (mode == MCOPY_ATOMIC_NORMAL)
 498                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 499                                                dst_addr, src_addr, page,
 500                                                wp_copy);
 501                 else
 502                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 503                                                  dst_vma, dst_addr);
 504         } else {
 505                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 506                                              dst_addr, src_addr,
 507                                              mode != MCOPY_ATOMIC_NORMAL,
 508                                              wp_copy, page);
 509         }
 510
 511         return err;
 512 }
 513
 514 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 515                                               unsigned long dst_start,
 516                                               unsigned long src_start,
 517                                               unsigned long len,
 518                                               enum mcopy_atomic_mode mcopy_mode,
 519                                               atomic_t *mmap_changing,
 520                                               __u64 mode)
 521 {
 522         struct vm_area_struct *dst_vma;
 523         ssize_t err;
 524         pmd_t *dst_pmd;
 525         unsigned long src_addr, dst_addr;
 526         long copied;
 527         struct page *page;
 528         bool wp_copy;
 529
 530         /*
 531          * Sanitize the command parameters:
 532          */
 533         BUG_ON(dst_start & ~PAGE_MASK);
 534         BUG_ON(len & ~PAGE_MASK);
 535
 536         /* Does the address range wrap, or is the span zero-sized? */
 537         BUG_ON(src_start + len <= src_start);
 538         BUG_ON(dst_start + len <= dst_start);
 539
 540         src_addr = src_start;
 541         dst_addr = dst_start;
 542         copied = 0;
 543         page = NULL;
 544 retry:
 545         mmap_read_lock(dst_mm);
 546
 547         /*
 548          * If memory mappings are changing because of non-cooperative
 549          * operation (e.g. mremap) running in parallel, bail out and
 550          * request the user to retry later
 551          */
 552         err = -EAGAIN;
 553         if (mmap_changing && atomic_read(mmap_changing))
 554                 goto out_unlock;
 555
 556         /*
 557          * Make sure the vma is not shared, that the dst range is
 558          * both valid and fully within a single existing vma.
 559          */
 560         err = -ENOENT;
 561         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 562         if (!dst_vma)
 563                 goto out_unlock;
 564
 565         err = -EINVAL;
 566         /*
 567          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 568          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 569          */
 570         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 571             dst_vma->vm_flags & VM_SHARED))
 572                 goto out_unlock;
 573
 574         /*
 575          * validate 'mode' now that we know the dst_vma: don't allow
 576          * a wrprotect copy if the userfaultfd didn't register as WP.
 577          */
 578         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 579         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 580                 goto out_unlock;
 581
 582         /*
 583          * If this is a HUGETLB vma, pass off to appropriate routine
 584          */
 585         if (is_vm_hugetlb_page(dst_vma))
 586                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 587                                                src_start, len, mcopy_mode,
 588                                                wp_copy);
 589
 590         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 591                 goto out_unlock;
 592         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 593                 goto out_unlock;
 594
 595         /*
 596          * Ensure the dst_vma has a anon_vma or this page
 597          * would get a NULL anon_vma when moved in the
 598          * dst_vma.
 599          */
 600         err = -ENOMEM;
 601         if (!(dst_vma->vm_flags & VM_SHARED) &&
 602             unlikely(anon_vma_prepare(dst_vma)))
 603                 goto out_unlock;
 604
 605         while (src_addr < src_start + len) {
 606                 pmd_t dst_pmdval;
 607
 608                 BUG_ON(dst_addr >= dst_start + len);
 609
 610                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 611                 if (unlikely(!dst_pmd)) {
 612                         err = -ENOMEM;
 613                         break;
 614                 }
 615
 616                 dst_pmdval = pmd_read_atomic(dst_pmd);
 617                 /*
 618                  * If the dst_pmd is mapped as THP don't
 619                  * override it and just be strict.
 620                  */
 621                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 622                         err = -EEXIST;
 623                         break;
 624                 }
 625                 if (unlikely(pmd_none(dst_pmdval)) &&
 626                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 627                         err = -ENOMEM;
 628                         break;
 629                 }
 630                 /* If an huge pmd materialized from under us fail */
 631                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 632                         err = -EFAULT;
 633                         break;
 634                 }
 635
 636                 BUG_ON(pmd_none(*dst_pmd));
 637                 BUG_ON(pmd_trans_huge(*dst_pmd));
 638
 639                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 640                                        src_addr, &page, mcopy_mode, wp_copy);
 641                 cond_resched();
 642
 643                 if (unlikely(err == -ENOENT)) {
 644                         void *page_kaddr;
 645
 646                         mmap_read_unlock(dst_mm);
 647                         BUG_ON(!page);
 648
 649                         page_kaddr = kmap(page);
 650                         err = copy_from_user(page_kaddr,
 651                                              (const void __user *) src_addr,
 652                                              PAGE_SIZE);
 653                         kunmap(page);
 654                         if (unlikely(err)) {
 655                                 err = -EFAULT;
 656                                 goto out;
 657                         }
 658                         flush_dcache_page(page);
 659                         goto retry;
 660                 } else
 661                         BUG_ON(page);
 662
 663                 if (!err) {
 664                         dst_addr += PAGE_SIZE;
 665                         src_addr += PAGE_SIZE;
 666                         copied += PAGE_SIZE;
 667
 668                         if (fatal_signal_pending(current))
 669                                 err = -EINTR;
 670                 }
 671                 if (err)
 672                         break;
 673         }
 674
 675 out_unlock:
 676         mmap_read_unlock(dst_mm);
 677 out:
 678         if (page)
 679                 put_page(page);
 680         BUG_ON(copied < 0);
 681         BUG_ON(err > 0);
 682         BUG_ON(!copied && !err);
 683         return copied ? copied : err;
 684 }
 685
 686 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 687                      unsigned long src_start, unsigned long len,
 688                      atomic_t *mmap_changing, __u64 mode)
 689 {
 690         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 691                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 692 }
 693
 694 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 695                        unsigned long len, atomic_t *mmap_changing)
 696 {
 697         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 698                               mmap_changing, 0);
 699 }
 700
 701 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 702                        unsigned long len, atomic_t *mmap_changing)
 703 {
 704         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 705                               mmap_changing, 0);
 706 }
 707
 708 void uffd_wp_range(struct mm_struct *dst_mm, struct vm_area_struct *dst_vma,
 709                    unsigned long start, unsigned long len, bool enable_wp)
 710 {
 711         struct mmu_gather tlb;
 712         pgprot_t newprot;
 713
 714         if (enable_wp)
 715                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 716         else
 717                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 718
 719         tlb_gather_mmu(&tlb, dst_mm);
 720         change_protection(&tlb, dst_vma, start, start + len, newprot,
 721                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 722         tlb_finish_mmu(&tlb);
 723 }
 724
 725 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 726                         unsigned long len, bool enable_wp,
 727                         atomic_t *mmap_changing)
 728 {
 729         struct vm_area_struct *dst_vma;
 730         unsigned long page_mask;
 731         int err;
 732
 733         /*
 734          * Sanitize the command parameters:
 735          */
 736         BUG_ON(start & ~PAGE_MASK);
 737         BUG_ON(len & ~PAGE_MASK);
 738
 739         /* Does the address range wrap, or is the span zero-sized? */
 740         BUG_ON(start + len <= start);
 741
 742         mmap_read_lock(dst_mm);
 743
 744         /*
 745          * If memory mappings are changing because of non-cooperative
 746          * operation (e.g. mremap) running in parallel, bail out and
 747          * request the user to retry later
 748          */
 749         err = -EAGAIN;
 750         if (mmap_changing && atomic_read(mmap_changing))
 751                 goto out_unlock;
 752
 753         err = -ENOENT;
 754         dst_vma = find_dst_vma(dst_mm, start, len);
 755
 756         if (!dst_vma)
 757                 goto out_unlock;
 758         if (!userfaultfd_wp(dst_vma))
 759                 goto out_unlock;
 760         if (!vma_can_userfault(dst_vma, dst_vma->vm_flags))
 761                 goto out_unlock;
 762
 763         if (is_vm_hugetlb_page(dst_vma)) {
 764                 err = -EINVAL;
 765                 page_mask = vma_kernel_pagesize(dst_vma) - 1;
 766                 if ((start & page_mask) || (len & page_mask))
 767                         goto out_unlock;
 768         }
 769
 770         uffd_wp_range(dst_mm, dst_vma, start, len, enable_wp);
 771
 772         err = 0;
 773 out_unlock:
 774         mmap_read_unlock(dst_mm);
 775         return err;
 776 }